Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Accurate rib fracture identification and classification are essential for treatment planning. However, existing datasets often lack fine-grained annotations, particularly regarding rib fracture characterization, type, and precise anatomical location on individual ribs. To address this, we introduce a novel rib fracture annotation protocol tailored for fracture classification. Further, we enhance fracture classification by leveraging cross-modal embeddings that bridge radiological images and clinical descriptions. Our approach employs hyperbolic embeddings to capture the hierarchical nature of fracture, mapping visual features and textual descriptions into a shared non-Euclidean manifold. This framework enables more nuanced similarity computations between imaging characteristics and clinical descriptions, accounting for the inherent hierarchical relationships in fracture taxonomy. Experimental results demonstrate that our approach outperforms existing methods across multiple classification tasks, with average recall improvements of 6% on the AirRib dataset and 17.5% on the public RibFrac dataset. The code and annotation files can be accessed at https://github.com/ribfracture123/classification

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1105_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ribfracture123/classification

Link to the Dataset(s)

N/A

BibTex

@InProceedings{PatShr_FineGrained_MICCAI2025,
        author = { Pate, Shripad and Farooq, Aiman and Datta, Suvrankar and Sheikh, Musadiq Aadil and Kumar, Atin and Mishra, Deepak},
        title = { { Fine-Grained Rib Fracture Diagnosis with Hyperbolic Embeddings: A Detailed Annotation Framework and Multi-Label Classification Model } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15974},
        month = {September},
        page = {221 -- 230}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a fine-grained annotation protocol for rib fractures by labeling key clinical attributes. This paper proposes a two-stage framework: first, a modified Faster R-CNN for fracture detection; second, a multi-head classification network to predict multiple fracture attributes. This paper presents a novel multimodal learning approach that projects both CT image features and clinical text descriptions into a shared hyperbolic space.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    Innovative Use of Hyperbolic Embeddings: Applying hyperbolic space to fuse image and text modalities is novel in medical imaging, enhancing the model’s ability to represent hierarchical fracture attributes. Annotation and Data Utilization: The detailed annotation of over 1,000 fractures provides rich information that improves fine-grained classification and benefits future research. Clinical Relevance: By directly computing clinical scores from model outputs, the work shows potential for practical use in treatment planning.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The detection module relies on established techniques, employing Faster R-CNN with ResNeXt without any improvements in the detection algorithm. Similarly, the multi-head classification network independently predicts fracture attributes—a strategy already explored in related work such as [1]. The hyperbolic embedding’s loss function and mapping formula are directly adopted from prior work, reducing the claimed originality [2].

    The compared baselines were not designed for the fine-grained task proposed, serving only as approximate references. The paper lacks comparisons with other widely used methods, such as UNet-based approaches [3,4].

    The limited data scale further weakens the impact. The AirRib dataset is annotated in detail, but the scale is relatively limited (there are only more than a hundred CT cases). The RibFrac dataset contains 660 cases, but only 50 cases were re-annotated and used for evaluation in this study [1]. This small dataset raises concerns about the statistical reliability and generalizability of the results when training a complex deep model with high-dimensional hyperbolic embeddings.

    [1] Yang J, Shi R, Jin L, et al. Deep rib fracture instance segmentation and classification from ct on the ribfrac challenge[J]. arXiv preprint arXiv:2402.09372, 2024. [2] Desai K, Nickel M, Rajpurohit T, et al. Hyperbolic image-text representations[C]//International Conference on Machine Learning. PMLR, 2023: 7694-7731. [3] Baumgartner M, Jäger P F, Isensee F, et al. nnDetection: a self-configuring method for medical object detection[C]//Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part V 24. Springer International Publishing, 2021: 530-539. [4] Isensee F, Jaeger P F, Kohl S A A, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation[J]. Nature methods, 2021, 18(2): 203-211.

  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    This paper contains erroneous repeated citations such as [21] and [22]

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Although the paper addresses an important clinical problem and attempts to integrate hyperbolic multimodal learning with fine-grained fracture annotation, significant concerns remain that warrant rejection. The overall framework relies heavily on established architectures without substantial improvements in detection or classification. The novelty of the hyperbolic embedding component is diluted by its direct adaptation from prior work. Furthermore, the evaluation uses a very limited dataset casting doubt on the generalizability and statistical reliability of the results. Given these limitations, I recommend rejection of the paper.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper “Fine-Grained Rib Fracture Diagnosis with Hyperbolic Embeddings: A Detailed Annotation Framework and Multi-Label Classification Model” makes three primary contributions to the field of automated rib fracture analysis from CT scans including fine-grained rib fracture annotation protocol, two-stage fine-grained rib fracture analysis pipeline, and hyperbolic multi-modal learning framework.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1. The authors introduce a novel and comprehensive annotation framework for rib fractures, addressing the lack of detailed, clinically relevant labels in existing datasets. Each fracture is annotated with multiple key clinical attributes includinga natomical location, displacement severity, morphological patterns, multiplicity, contribution to complex injuries, rib side and number.

    2. The paper proposes a two-stage deep learning pipeline: Detection Stage: A modified Faster R-CNN detects potential rib fractures in CT scans, using customized anchor boxes and a post-processing algorithm to ensure spatial and temporal coherence across slices. Classification Stage: A novel multi-head classification network characterizes each detected fracture along multiple clinically relevant dimensions, using a 3D ResNet-50 backbone for feature extraction.

    3. The core methodological innovation is the use of hyperbolic embeddings to jointly model radiological images and clinical text descriptions in a shared non-Euclidean space. Both visual features (from CT patches) and textual features (from ClinicalBERT-encoded descriptions) are projected into hyperbolic space, capturing the hierarchical relationships inherent in fracture taxonomy.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. The paper only describes the total size, source, and partitioning of the dataset, without presenting detailed statistics or visualizations of the dataset distribution. This represents a clear information gap in terms of fairness evaluation, class imbalance analysis, and reproducibility.
    2. The 2D detection approach with slice tracking (IoU 0.1%, 4-slice minimum) may struggle with some fracture types such as markedly displaced fracture, differentiating rib fractures from costochondral junctions. Compared to native 3D detectors, this pipeline risks losing spatial context critical for displacement assessment.
    3. Critical parameters like hyperbolic curvature (κ=1.0) and entailment weight (λ=0.2) are fixed without ablation studies. The impact of these choices on hierarchical relationship modeling remains unquantified.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The fine-grained rib fracture annotation protocol clearly demonstrates the clinical need to characterize rib fracture features. However, the absence of detailed dataset characteristics and distribution undermines the credibility of the experimental results.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    In this study, the authors first outlined three primary contributions, which can be further elaborated upon based on the content of the manuscript as follows: [1] Establishing a fine-grained annotation protocol for rib fractures. While most previous studies focused solely on binary classification, this study introduced a multi-label framework by considering four clinically relevant attributes. [2] Developing a two-stage pipeline for fine-grained rib fracture analysis. The authors designed a network that performs both detection and multi-class classification. [3] Proposing a hyperbolic multi-modal learning framework to integrate imaging and clinical text. They constructed a classification model that incorporates image features and text embeddings, reflecting a hierarchical structure.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    [1] The study did not merely aim to detect the presence or absence of fractures, but rather classified clinically meaningful attributes that are considered by radiologists during diagnosis, and incorporated these classifications into the computation of the RibScore metric. [2] By utilizing hyperbolic space embeddings, the study effectively captured the hierarchical semantic structure of rib fracture classifications and designed a loss function that reflects this structure. [3] The dataset was manually annotated, and the labeling information was made publicly available to ensure reproducibility.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    [1] Since the dataset (AirRib) primarily consisted of data from a single institution, there were inherent limitations in terms of generalizability. [2] Although ClinicalBERT was used for text embedding, the study lacked a detailed explanation of this component, making it difficult to determine which clinical terms had a significant influence on the classification. [3] Rib fracture detection was performed in 2D, while the subsequent processes were carried out in 3D, which may have introduced inconsistencies. A comparison with a fully 3D-based detection approach is necessary to evaluate potential performance differences.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The study presents novelty in its ability to perform multi-classification and its use of hyperbolic embeddings. However, more detailed explanations are needed regarding the integration of text embeddings and the combination of 2D and 3D processes. Furthermore, considering the characteristics of the dataset, it may be difficult to generalize the results.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank reviewers R1, R3, and R4 for their thoughtful feedback and constructive suggestions. We have carefully addressed each comment as detailed below. [R1W1] Our AirRib dataset has 1,014 fractures across 103 CT scans with balanced distribution: 50.29% left side, 49.70% right side; rib coverage of T1: 7.29%, T2-T8: 77.12%, T9-T12: 15.59%; locations include Anterior: 15.98%, Lateral: 41.91%, Posterior: 42.11%. Displacement patterns show Buckle: 8.28%, Non-displaced: 32.64%, Displaced: 28.70%, Severely Displaced: 30.37%. Fracture types include Linear Oblique: 65.78%, Buckle: 8.28%, Complex: 17.55%, Linear Transverse: 8.38%, with patterns of Not Applicable: 54.04%, Segmental: 13.21%, Flail: 32.7%. For RibFrac (473 fractures across 50 scans), complete details are in the published dataset. [R1W2] The threshold values (IoU 0.1%, 4-slice minimum) were chosen through testing to optimize detection across fracture types. These settings balance sensitivity and specificity, capturing most fractures while reducing false positives. Our tests show these parameters work well even for displaced fractures. The pipeline maintains spatial context needed for displacement assessment, with performance comparable to fully 3D approaches. [R1W3] We chose hyperbolic curvature κ=1.0 and entailment weight λ=0.2 based on testing and prior research. These values consistently gave the best results and match established values in previous hierarchical modeling studies. [R3W1] While we use established components (Faster R-CNN with ResNeXt for detection and multi-head classification), our innovation lies in their novel integration for rib fracture diagnosis. We are the first to apply hyperbolic embeddings to medical imaging for hierarchical fracture classification. Though we build upon foundations from Desai et al., we’ve adapted the hyperbolic space specifically for radiological findings by introducing domain-specific constraints that capture clinical relationships in fracture taxonomy. This adaptation enables direct computation of clinical metrics (RibScore) from our model outputs - a capability not present in previous work. The effectiveness of our approach is demonstrated by the 6-17.5% recall improvement over existing methods, validating our contribution to rib fracture analysis. [R3W2] We chose baseline methods (CT, CS, UCI) based on their direct relevance to rib fracture classification. UNet methods are good for segmentation but would need significant changes to handle our multi-label classification across four clinical dimensions. Our baselines represent current best approaches specifically for rib fractures. We will include comparisons with adapted UNet methods in future work, especially for the detection stage. [R3W3, R4W1] We acknowledge that our study is limited by dataset size. While AirRib has detailed annotations, having only 103 CT cases limits reliability. We have annotated the 50 cases from the RibFrac dataset according to our classification method. We plan to use larger multi-institutional datasets in future work to further validate our model. [R4W2] We used ClinicalBERT to embed specific clinical terms important for classification, including descriptions like “non-displaced fracture,” “oblique,” and locations like “posterior aspect of the right side.” These terms capture key features radiologists use when examining rib fractures. By focusing on these specific terms rather than full reports, our model targets the most important diagnostic information. This approach improves accuracy while keeping the model interpretable, as each term directly relates to clinical fracture features. [R4W3] We performed detection in 2D for computational efficiency, but conducted all evaluation in 3D. After detecting 2D bounding boxes, we generated 3D volumes and performed all subsequent operations within this 3D context. This hybrid approach preserves crucial spatial relationships while requiring significantly less computing power than full 3D detection meth



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A



back to top